acceptodds
Under review as a conference paper at ICLR 2027

OffRoad-VLA: Terrain‑Aware Action Tokens and Affordance Reasoning for Off-Road VLA Navigation

Abstract

Vision-Language-Action (VLA) models have recently gained significant attention in autonomous driving for their ability to leverage world knowledge for navigation. However, existing VLA approaches primarily target on-road environments, where physical actions assume structured road surfaces and traffic semantics. We present OffRoad-VLA, an end-to-end autoregressive VLA framework that unifies terrain-aware reasoning and physical action generation for off-road navigation. We introduce a terrain-aware physical action codebook that jointly discretizes vehicle motion with local terrain properties, including slope, roughness, and traversability, enabling the model to generate kinematically feasible and terrain-adaptive trajectories. To support efficient reasoning, we employ dual thinking modes that produce direct actions for simple scenes and adaptive thinking using chain-of-thought (CoT) reasoning for complex scenario, explicitly connecting terrain conditions and obstacles to driving actions. We further apply Group Relative Policy Optimization (GRPO) with off-road safety rewards to improve planning while penalizing traversability violations, unsafe terrain interactions, encouraging progress, and smooth motion. Experiments on RELLIS-3D and TartanDrive, together with zero-shot evaluation on the unseen GOOSE dataset, demonstrate improved long-horizon trajectory prediction, terrain classification, and safety over existing approaches. Closed-loop evaluation in the MAVS physics-based simulator further demonstrates robust real-time navigation across diverse off-road conditions.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Vision Language Action Model, Terrain-Aware Action Tokenization, Reinforcement Finetuning, End-to-end Off-road Autonomous Driving

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_gBMcOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This work introduces OffRoad‑VLA, an end‑to‑end autoregressive vision‑language‑action framework targeted at unstructured off‑road navigation. The paper designs a terrain‑aware physical action codebook which jointly discretizes vehicle kinematics together with local terrain properties including slope, roughness and traversability. It adopts dual‑mode thinking: direct‑action generation for simple terrain scenes and affordance‑centric chain‑of‑thought (CoT) reasoning for complex hazardous conditions. A two‑stage training pipeline is adopted, supervised adaptation (SA) learns from expert trajectories and distilled terrain reasoning generated by a large teacher VL‑LM, safety‑oriented post‑training reinforcement learning (PRL) with GRPO further optimizes planning while discouraging unsafe terrain interactions. Evaluations are conducted on RELLIS‑3D and TartanDrive, plus zero‑shot tests on the unseen GOOSE dataset. Closed‑loop validation is performed within the physics‑based MAVS simulator.

Critical strengths
  1. This paper focuses on the challenging off‑road domain where slope, roughness and traversability dominate safety, filling a clear research gap for VLA‑based outdoor navigation.

  2. The proposed work embeds terrain attributes directly into discrete action tokens. This enables the VLM backbone to reason about motion and terrain safety jointly, rather than relying on external standalone terrain planners.

  3. The model can dynamically choose between direct‑action output for easy cases and detailed terrain‑grounded CoT for complex risky scenes. Automated teacher‑model distillation reduces heavy manual annotation cost for off‑road reasoning data.

  4. The authors carry out in‑distribution open‑loop tests, zero‑shot generalization on fully unseen GOOSE dataset, plus closed‑loop physical simulation on MAVS.

Critical weaknesses
  1. Many components are assembled from existing prior arts instead of original innovations. The overall pipeline appears to be a recombination of well‑established building blocks: action tokenization for VLA robot systems, chain‑of‑thought distillation from large teacher models, and GRPO reinforcement learning for vision‑language models. The paper repurposes these known techniques for off‑road scenarios, but few fundamentally new algorithmic insights are introduced.

  2. CoT labels are automatically generated by a 72B VL‑LM teacher model. Even with human sanity checks, annotation noise or logical flaws in distilled reasoning may propagate to the student model, and the robustness of reasoning under highly novel terrain is not thoroughly analyzed.

  3. The paper presents final combined performance of the full reward function, but lacks controlled ablations to isolate how each reward term (traversal penalty, slope penalty, reasoning‑length penalty etc.) independently contributes to the final model behaviour.

  4. The idea of switching between direct output and CoT reasoning conditioned on scenario complexity has already been explored in multiple embodied‑AI and reasoning‑VLA literature. This work merely adapts this existing paradigm to off‑road terrain inputs, without proposing new reasoning mechanisms tailored for terrain‑driven decision‑making.

  5. It is difficult to disentangle whether performance gains come from the proposed terrain‑aware design, or simply stem from stronger baseline capabilities brought by the large‑capacity Qwen2.5‑VL backbone, distillation and GRPO fine‑tuning. The paper does not set up sufficient controlled baselines that isolate the exact contribution of each supposedly novel module.

  6. The abstract claims robust real-time navigation, but the reported evidence does not appear to support this claim. Table 4 reports results for the SA model rather than PRL, while MAVS pauses physics during inference, meaning that the reported 11.1 s (SA) or 1.6 s (PRL) inference times do not directly affect the simulated vehicle dynamics. Moreover, the conclusion itself acknowledges that the 2 Hz onboard setting is not sufficiently supported. These inconsistencies make the claim of robust real-time navigation difficult to substantiate.

  7. Slope, roughness, and traversability are baked into action tokens, reused as GRPO penalties, and reused as TVR/collision. Appendix E.3 still feeds measured slope/traversability as input. That does not show vision-based terrain understanding.

  8. Several statements regarding the reported improvements appear inconsistent with the numerical results. For example, ADE-5s decreases from 2.45 to 2.15, corresponding to a 12.24% reduction, which does not support a claim that the error is nearly halved. Moreover, the short-horizon error reduction is approximately 9%, while FDE-5s actually increases from 5.43 to 5.69. On RELLIS, the reported 0.75 m drop appears to refer to the absolute FDE level rather than the 0.06 m SA-to-PRL improvement. Similarly, the TartanDrive TCRC is reported as 5.27%, which should be reconciled with the corresponding claims in the text. These discrepancies undermine the precision and credibility of the paper's quantitative claims.

Questions for authors
  1. What fundamentally new algorithmic or reasoning mechanism does the paper introduce beyond combining existing VLA, CoT distillation, and GRPO techniques?

  2. Are the gains genuinely attributable to terrain-aware reasoning, or primarily to the stronger backbone, teacher distillation, and RL fine-tuning?

  3. Does the system truly perform camera-based terrain understanding and robust real-time navigation, given the use of explicit terrain information and the limitations of the reported simulation protocol?

  4. Are the ablations, baselines, and quantitative interpretations sufficiently controlled and accurate to support the paper's main conclusions?

3 comments

We sincerely thank the reviewer for the comment and suggestions. We have tried to address most of the suggestions. W, Q means address to the weak points, and questions.

W1, W4/Q1: We thank the reviewer for this concern. We agree that the supervised adaption approach and GRPO-based fine-tuning is well-studied in the driving domain. However, these approaches primarily target structured on-road diving (such as traffic analysis, left-right turning) and do not explicitly address off-road conditions (such as adaption with different terrain conditions). We already acknowledged those in Section 2 (lines 155-158)\color{blue}{\text{Section 2 (lines 155-158)}}. Our main contribution is to develop an off-road VLA capable of encoding diverse terrain properties into its action representations, enabling terrain-aware reasoning and navigation. We tried to address this and shown why it does not transfer directly:

  1. Existing tokenized driving VLAs use purely kinematic codebooks. Our codebook jointly clusters motion with traversability class, slope, IMU roughness and a velocity-conditioned cost (as mentioned in Section 3.2, lines 226-237, Eq. 14), so every token carries a terrain attributes. This is what allows the traversability and slope rewards in PRL to be computed directly on the generated tokens, without any terrain map at inference (discussed in Appendix D.1, lines 1282-1292\color{blue}{\text{discussed in Appendix D.1, lines 1282-1292}}).

  2. Secondly, on-road conditions expect a single shared codebook, although it fails in off-road. In off-road codebook construction comes from platform dynamics. Therefore, direct transfer of codebook construction approach from on-road to off-road is not feasible. We have already given justification in Appendix A.3.1, lines 1050-1056\color{blue}{\text{Appendix A.3.1, lines 1050-1056}}.

  3. Moreover we did the affordance-centric COT reasoing. Which is a crucial part of our proposed approach and is necessary for off-road navigation. We have discussed its advantage in Appendix B\color{blue}{\text{Appendix B}}.

W5/Q2: Table 5b isolates terrain-aware tokens while keeping the backbone, data and training identical. Terrain tokens alone reduce FDE-5s from 0.94 to 0.81 m and raise terrain classification from 70.32% to 79.76% and SS from 0.73 to 0.85. SA vs. PRL isolates the effect of RL (as discussed in Tables 1-2\color{blue}{\text{as discussed in Tables 1-2}}). In addition, Alpamayo also combines a VLM, reasoning and RL, however it reaches 49.74% collisions on RELLIS-3D versus 1.32% for our SA model (see Table 1\color{blue}{\text{see Table 1}}). This shows that a strong backbone with CoT and RL is not sufficient without terrain-aware design for off-road navigation task.

W2: The teacher receives the ground-truth action as a hint, so its reasoning is anchored to the correct decision (as discussed in Appendix B.1\color{blue}{\text{as discussed in Appendix B.1}}). Moreover, human annotators checked 1,000 samples under a strict all-correct criterion (getting 83.47% accuracy), and failing samples were removed. Moreover, PRL rewards are computed on actions, not on reasoning text, so flawed reasoning that leads to unsafe actions is penalized. On unseen GOOSE terrain, the model keeps 0% collisions and TVR and produces coherent reasoning.

W3: All reward terms except r_CoT depend only on action tokens, so the 94% reduction in reasoning tokens after PRL (as mentioned in Appendix F.4\color{blue}{\text{as mentioned in Appendix F.4}}) is attributable to r_CoT. r_trav and r_slope target exactly the terrain classes counted by TVR, which drops from 0.37% to 0.03% (on RELLIS-3D) and from 0.13% to 0.07% (on TartanDrive).

W6/Q3: We clarify that the 1.6 s inference latency and the 2 Hz replanning rate measure different aspects. The 1.6 s represents the wall-clock inference latency of the PRL model on TartanDrive dataset, and measured using unoptimized decoding on a single GPU (also mentioned in Appendix F.4\color{blue}{\text{Appendix F.4}}). In contrast, 2 Hz is the replanning rate in simulated time under the MAVS evaluation protocol. Moreover, inference latency also varies across datasets because it depends on the number of generated tokens. On TartanDrive, SA requires 11.1 s to generate 554.7 tokens, whereas PRL requires only 1.6 s for 45.0 tokens. We pause the simulator during inference so that closed-loop results reflect planning quality.

As stated in the Conclusion, these results do not establish real-time onboard execution. However, we will replace "real-time" in the Abstract of the revised manuscript with "closed-loop navigation at a 2 Hz replanning rate in simulation."

W7/Q3: Terrain labels are used only during training, to build the codebook and compute rewards. At inference, the model receives camera images, ego state (v, a, ω, action history, this is clearly mentioned in Section 3.1\color{blue}{\text{this is clearly mentioned in Section 3.1}}) and the instruction. However, Appendix E.3’s mention of “terrain priors” is inaccurate and will be corrected in the revision. TVR and collision are evaluated against ground-truth terrain under the predicted waypoints, not against the model’s own tokens. On GOOSE, whose labels are never used in training, the model still achieves 0% collisions and TVR, indicating camera-based terrain understanding.

W8/Q4: We thank the reviewer and will correct these statements. On TartanDrive, ADE improves by 8.6%, 9.8% and 12.2% at 1/3/5 s (not “nearly halved”), and FDE-5s increases from 5.43 to 5.69 m. On RELLIS-3D, FDE decreases by 0.06 m, to 0.75 m. TartanDrive TCRC (5.27%) is low in absolute terms because of the strict 1 m goal criterion over a 16 m horizon, but it remains well above all baselines (≤0.67%). We will also fix the typo in Table 9 (−26%, not −2696%).

We thank the reviewer again for the careful review, which has helped us sharpen our contributions and correct the presentation. We hope our responses address the concerns raised.

My final recommendation is acceptance as the rebuttal clearly addresses my major concerns. Iwill raise my score.

Reviewer_dS6nOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes OffRoad-VLA, an end-to-end vision-language-action framework for off-road navigation that combines terrain-aware action tokenization, affordance-centric chain-of-thought reasoning, and safety-aware reinforcement learning. The method uses a terrain-conditioned action codebook, adaptive reasoning for complex scenarios, and GRPO-based post-training to improve trajectory generation and safety. Experiments on RELLIS-3D, TartanDrive, and the unseen GOOSE dataset report improvements in planning accuracy and terrain safety, while closed-loop MAVS simulations demonstrate competitive navigation performance.

Critical strengths
  1. The paper addresses an important and relatively underexplored problem in VLA-based autonomous navigation, where terrain properties and vehicle dynamics significantly affect safe trajectory generation.
  2. The integration of terrain-aware action tokens with affordance-based reasoning is interesting, and the evaluation covers multiple datasets, zero-shot generalization, and closed-loop simulation. The controlled comparison against kinematics-only tokenization provides useful evidence for the proposed representation.
Critical weaknesses
  1. The real-time navigation claims are not convincingly supported. The reported inference latency is 1.6 seconds, substantially exceeding the 0.5-second budget required for 2 Hz operation. Moreover, the MAVS simulator is paused during inference, meaning the closed-loop experiments do not capture the effects of inference delays on vehicle dynamics. A separately trained MAVS model is also used, making it difficult to establish how well the original policy transfers to closed-loop deployment.
  2. The safety and generalization claims appear stronger than the experimental evidence supports. The GOOSE evaluation contains only 70 test samples, and despite reporting zero collisions and traversability violations, the TartanDrive-trained model still exhibits high miss rates (83% at 2 m and 55.6% at 4 m). Furthermore, the closed-loop driving score is not the best among compared methods, contrary to the broader performance claims.
  3. The independent contributions of several proposed components remain insufficiently validated. While terrain-aware tokenization is ablated, the effects of adaptive CoT reasoning, individual safety reward components, and reasoning-efficiency optimization are not systematically isolated. Some reported improvements are also inconsistent, including a worsening of TartanDrive FDE after reinforcement learning.
Questions for authors
  1. How can the proposed method support real-time 2 Hz navigation when inference requires 1.6 seconds? Can the authors evaluate closed-loop performance without pausing the simulator?
  2. Why is a separately trained MAVS model required for closed-loop evaluation, and does the evaluation use the SA or PRL policy? Can the authors report closed-loop results that isolate the benefit of PRL?
  3. Can the authors provide independent ablations for adaptive reasoning, CoT supervision, and individual GRPO reward components to establish their contributions?
  4. How reliable are the zero-shot safety conclusions given the small GOOSE test set and relatively high trajectory miss rates? Can the authors clarify the reported performance inconsistencies and provide uncertainty estimates?
2 comments

We sincerely thank the reviewer for the comment and suggestions. We have tried to address most of the suggestions. W, Q means address to the weak points, and questions.

W1/Q1-Q2: We clarify that the 1.6 s inference latency and the 2 Hz replanning rate measure different aspects. The 1.6 s represents the wall-clock inference latency of the PRL model on TartanDrive dataset, and measured using unoptimized decoding on a single GPU (also mentioned in App. F.4\color{blue}{\text{App. F.4}}). In contrast, 2 Hz is the replanning rate in simulated time under the MAVS evaluation protocol. Moreover, inference latency also varies across datasets because it depends on the number of generated tokens. On TartanDrive, SA requires 11.1 s to generate 554.7 tokens, whereas PRL requires only 1.6 s for 45.0 tokens. We pause the simulator during inference so that closed-loop results reflect planning quality.

As stated in the Conclusion, these results do not establish real-time onboard execution. However, we will replace "real-time" in the Abstract of the revised manuscript with "closed-loop navigation at a 2 Hz replanning rate in simulation."

W2/Q2: A separate MAVS environment-specific model is necessary, just as we train separate models for RELLIS-3D and TartanDrive. This is because the action codebook is fitted to each platform's motion characteristics (we already described it in detailed in App. A.3.1\color{blue}{\text{App. A.3.1}}). MAVS has a different vehicle, camera configuration, and speed range. Consequently, a codebook fitted to another platform cannot adequately represent MAVS kinematics. Therefore, a separately trained MAVS model is required for valid closed-loop evaluation.

W2/Q4: Col. and TVR measure whether the predicted trajectory stays on traversable terrain, whereas MR measures endpoint deviation from a single expert trajectory over 5 s on an unseen platform. Therefore, a safe trajectory may still have a higher MR if it differs from the expert endpoint. Nevertheless, MR2 and MR4 remain high across all methods, while OffRoad-VLA achieves the lowest endpoint error.

Regarding DS, OFF-EMMA achieves a marginally higher score (0.93 vs. 0.91), although drives at only 0.82 m/s. In contrast, OffRoad-VLA achieves a much higher average speed of 4.31 m/s (approximately 5× faster) and a higher success rate (84% vs. 81.7%). Since DS does not explicitly account for traversal speed, slower and more conservative driving can achieve fewer infraction penalties. Thus, DS alone does not fully capture the trade-off between driving safety and efficiency.

W3/Q3: We thank the reviewer for this comment. Although, existing experiments already demonstrate the contribution of each component, as summarized below.

Reasoning efficiency: In Eq. 15, all reward terms depend on action tokens except (r_CoT), which explicitly regulates reasoning length. On TartanDrive, PRL reduces reasoning tokens from 539.7 to 30.0 (by 94.4%) and inference latency from 11.1 s to 1.6 s. This reduction does not compromise planning performance. On RELLIS-3D, route completion improves from 81.58% to 84.74% (shown in Fig. 3a\color{blue}{\text{Fig. 3a}}).

Adaptive reasoning: During SA, the model is trained with CoT on approximately 30% of samples and direct action prediction on the remaining samples (discussed in Sec. 3.3\color{blue}{\text{Sec. 3.3}}). This allows the model to learn when explicit reasoning is necessary. Fig. 3c illustrates how PRL enables the model to switch from reasoning to direct action prediction on simple terrain.

Safety rewards: The (r_trav) and (r_slope) terms penalize unsafe terrain. After PRL, TVR decreases from 0.37% to 0.03% on RELLIS-3D and from 0.13% to 0.07% on TartanDrive. Collision rates also decrease from 1.32% to 0.26% and from 1.28% to 0.67%, respectively (detailed in Tables 1-2\color{blue}{\text{Tables 1-2}}). Meanwhile, terrain classification accuracy improves from 79.76% to 84.11% and from 85.62% to 89.85% (discussed in Tables 11-12\color{blue}{\text{Tables 11-12}}). Additional experiments show that varying the SA terrain weight (\lambda_t) between 1.0 and 2.0 changes TVR by less than 3%, while GRPO remains stable across different group sizes (see visual and explanation result in Fig. 3b\color{blue}{\text{visual and explanation result in Fig. 3b}}). Also, terrain-aware tokenization is evaluated separately in Table 5b\color{blue}{\text{Table 5b}}. Together, these results demonstrate the effectiveness of the proposed components.

TartanDrive FDE: After PRL, FDE increases slightly from 5.43 m to 5.69 m (4.8%). However, ADE improves across all prediction horizons (8.6%, 9.8%, and 12.2%), while collision rates and TVR decrease substantially. This trade-off arises because the accuracy reward optimizes average trajectory error (ADE), whereas safety rewards encourage trajectories to avoid unsafe terrain. At higher speeds, these safety-driven adjustments increase the final displacement error. Nevertheless, our FDE remains lower than all baselines. On RELLIS-3D, FDE also improves from 0.81 m to 0.75 m.

We sincerely thank the reviewer for the constructive feedback, which has helped us clarify our claims and strengthen the paper, and we hope our responses have addressed the concerns raised.

Reviewer_oNQVOpenReview ↗
Rating
2(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The authors proposed OffRoad-VLA, which fine-tunes Qwen2.5-VL-3B for off-road driving with an action codebook that jointly clusters motion and terrain attributes. It adds teacher-distilled affordance CoT with direct/thinking modes and GRPO with hazard-aware rewards. It is evaluated on RELLIS-3D, TartanDrive, zero-shot GOOSE, and MAVS closed-loop.

Critical strengths
  1. The idea of off-road is interesting, and attaching terrain attributes to action tokens is reasonable.
Critical weaknesses

Weaknesses:

  1. Most of the components are well-studied and well-used in the driving domain, i.e., SFT, duel system, GRPO. In my opinion, it is more like an engineering implementation and lacks novelty.

  2. Following the first point, I am curious whether you have tried directly using other existing autonomous driving VLA models to address those datasets? According to Table 1, only Alpamayo belongs to this category, and in addition, it is not fair to compare Alpamayo with OffRoad-VLA since Alpamayo doesn't have any components for encoding off-road features.

  3. In addition to Alpamayo, I am not convinced about the results in Table 1, as other baselines don't have off-road features, it is not a fair comparison.

Minors:

  • Line 375 0.0.07%
  • Offroad-VLA or OffRoad-VLA?
Questions for authors

Please see the Weaknesses section. The main concern is the fair comparison. It is not fair to raise a new question/task and use a proposed tailored method to compare it with existing baselines. It is obvious that the proposed one will be the best.

1 comment

We sincerely thank the reviewer for the thoughtful comments and suggestions. We have tried to address them.

W1: We thank the reviewer for this concern. We agree that the supervised adaption approach and GRPO-based fine-tuning is well-studied in the driving domain. However, these approaches primarily target structured on-road diving (such as traffic analysis, left-right turning) and do not explicitly address off-road conditions (such as adaption with different terrain conditions). We already acknowledged those in Section 2 (lines 155-158)\color{blue}{\text{Section 2 (lines 155-158)}}. Our main contribution is to develop an off-road VLA capable of encoding diverse terrain properties into its action representations, enabling terrain-aware reasoning and navigation. We tried to address this and shown why it does not transfer directly:

  1. Existing tokenized driving VLAs use purely kinematic codebooks. Our codebook jointly clusters motion with traversability class, slope, IMU roughness and a velocity-conditioned cost (as mentioned in Section 3.2, lines 226-237, Eq. 14), so every token carries a terrain attributes. This is what allows the traversability and slope rewards in PRL to be computed directly on the generated tokens, without any terrain map at inference (discussed in Appendix D.1, lines 1282-1292\color{blue}{\text{discussed in Appendix D.1, lines 1282-1292}}).

  2. Secondly, on-road conditions expect a single shared codebook, although it fails in off-road. In off-road codebook construction comes from platform dynamics. Therefore, direct transfer of codebook construction approach from on-road to off-road is not feasible. We have already given justification in Appendix A.3.1, lines 1050-1056\color{blue}{\text{Appendix A.3.1, lines 1050-1056}}.

  3. Moreover we did the affordance-centric COT reasoing. Which is a crucial part of our proposed approach and is necessary for off-road navigation. We have discussed its advantage in Appendix B\color{blue}{\text{Appendix B}}.

W2: We agree that Alpamayo is not a like-for-like baseline. However, it aims to generalize autonomous driving by bridging reasoning and action prediction, particularly for long-tail scenarios. Therefore, we include Alpamayo for both in-domain and cross-domain comparisons in off-road settings. The results show that it performs substantially worse than our proposed OffRoad-VLA in these settings.

W3: We would like to clarify that OFF-EMMA and VAMOS are designed for outdoor environments, where off-road conditions are common. Furthermore, GNM, ViNT, and NoMaD are pretrained on multi-robot navigation datasets that include off-road driving. To the best of our knowledge, however, there are currently no dedicated off-road driving VLA models available for direct comparison with our proposed approach.

Comments on the paper

2 comments
Area_Chair_vvMLOpenReview ↗

Reminder: Author Response and Reviewer Discussion

Dear Reviewers,

We are halfway through the discussion period. Please review the authors’ responses, address any remaining concerns, and engage in the discussion where appropriate. If your assessment remains unchanged, please briefly acknowledge that you have considered the authors’ responses.

Thank you for your time and contributions.

Dear Reviewers,

We sincerely thank you for your time and constructive feedback on our submission. We have carefully considered every comment and addressed each one in our individual responses.

In preparing these responses, we have aimed to point you to evidence already present in the submission, including the ablations, analyses and appendix results that support the contribution of each component. We also appreciate you pointing out claims that were stated more strongly than our results support. These overclaimed statements, particularly those on real-time navigation and on the reported improvements, have been revised and corrected in the updated manuscript to match the experimental evidence.

Due to overlapping submission deadline during the rebuttal period, our participation in the discussion was more limited than usual. To make up for this, we have provided detailed responses to all your comments, and we hope they help you follow the updates in the revised manuscript.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.